Building upon existing methods for three-dimensional geometric rule recognition， this paper proposes a sealed-cavity notch feature recognition algorithm based on an attributed adjacency graph （AAG）. The algorithm first constructs a face-edge AAG to achieve a unified representation of geometric and topological information， thereby providing a more systematic and intuitive description framework for complex structural features. On this basis， the proposed method integrates a notch template library with an improved VF2 subgraph isomorphism algorithm to automatically match and identify potential notch regions within the target model. To enhance computational efficiency， a node constraint mechanism and heuristic ordering strategy are introduced during the search process， effectively mitigating the state-space explosion problem and significantly reducing computational complexity. Experimental results demonstrate that the proposed approach outperforms traditional rule-based methods in the automatic recognition of multiple types of notch features， exhibiting superior robustness， accuracy， and generality. This work provides an efficient and extensible solution for automatic detection and feature recognition in three-dimensional geometric modeling.